01What you are watching
Authenticate
The agent signs in through the portal’s own login form. There is no API to call. Credentials are injected at runtime and never written to disk or into the trace.
Survive the interstitial
The portal interrupts login with a password-expiry notice. A naive script treats that as a failed login. The agent recognises it, dismisses it, flags it for the operator and carries on.
Extract
The account register is read page by page, with the row count reconciled at each step so a silently short page is caught rather than averaged away.
Absorb the rate limit
Pushing harder earns an HTTP 429. The agent treats throttling as expected behaviour rather than an error, engages adaptive backoff, and resumes at a reduced rate.
Fail loudly, not silently
An upstream node returns 502. The circuit breaker opens and an alert is raised. A dashboard that stays green while data goes stale is the worse outcome, and this is the design decision that prevents it.
Reconcile, escalate, audit
Billed amounts are matched against contract rates, one ambiguous record is escalated to a human with its evidence attached, and the run closes in dry-run mode: zero writes, full trace on disk.
02What this does and does not prove
Read this before you draw conclusions
- It proves the mechanism. An agent can authenticate into a portal with no API, work through it, extract structured data and reconcile it — unattended, with a trace.
- It does not prove it will work on your system. Every portal fails differently. That is exactly what discovery is for, and why we capture real traffic before quoting.
- Nothing here is a client system. We do not record client environments, and we do not access any system without written authorisation.
- Dry run means dry run. Every write is suppressed in this recording. The diff is produced for review rather than applied.
03Why it holds up in production
The difficult part of this work is not the happy path you have just watched — it is everything around it. Portals change without notice, so the agent detects structural change and raises an alert rather than continuing against a page it no longer understands. Rate limits are respected with adaptive backoff rather than fought. Credentials are injected at runtime and never written to disk. Every run is logged, replayable and reversible.
Silent failure is the worst outcome in this kind of automation: a dashboard that stays green while the data quietly goes stale. Most of the engineering exists to make failure loud.
See whether it works on your system
A fixed-price discovery establishes whether your specific workflow can be automated — including when the honest answer is that it cannot, or should not be.
Start a fixed-price discovery →